Real-time, machine learning-enhanced position recognition for sub-half-wavelength precision laser nanofabrication
image: (a) Schematic of the laser nanofabrication system. (b) Simulated and experimentally captured images. N-defocus indicates the focal point below the sample surface, In-focus indicates the sample at the optimal focal position for clear laser pattern delineation on the fabrication material, and P-defocus indicates the focal spot above the sample surface.
Credit: Hao Zhang, Jinchuan Zheng, Guiyuan Cao, Han Lin, Baohua Jia
A camera already installed for routine viewing is now enough to keep femtosecond-laser fabrication sharply focused. Writing in Engineering, a team from Swinburne University of Technology, Shenzhen University and RMIT University shows that a three-layer neural network can read the shape of the focal spot on graphene oxide and report the axial distance between beam and surface with an accuracy of 257 nm—about half the 800 nm fabrication wavelength.
The group trained the network on 1 516 live images recorded at 23 frames per second while the sample, tilted 7°, was scanned under an Olympus 50× oil-immersion objective. Each 5-megapixel frame was resized to 205 × 205 pixels and cropped to a 51 × 51-pixel region that carries the spot intensity. Linear interpolation supplied the missing micrometre-scale labels, and an 85 %-training/15 %-testing split delivered the lowest root mean square error (RMSE).
Compared with rational-quadratic Gaussian-process regression, kernel-approximation least squares and quadratic support-vector machines, the trilayer network reached the smallest training RMSE (0.3427 µm) and testing RMSE (0.2566 µm) while keeping prediction speed at 1 000 observations per second. Detection uncertainty stayed within ±0.4 µm across the ±2 µm in-focus window that matters for writing 200-nm-thick GO lenses.
Scanning-electron micrographs taken after spiral-pattern exposure show that structures written without compensation fade at the edges where the beam wandered out of focus, whereas the network-corrected trajectory yields uniform, high-conductivity lines. Because no extra optics or sensors are required, the software-only upgrade can be retro-fitted to existing laser nanofabrication platforms that already carry a CMOS camera.
The paper “Real-Time Machine Learning-Based Position Recognition in Laser Nanofabrication with Sub-Half-Wavelength Precision,” is authored by Hao Zhang, Jinchuan Zheng, Guiyuan Cao, Han Lin, Baohua Jia. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.03.037. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.